{"id":"47651b61-ff98-41cb-806e-b7efbd08e20b","arxiv_id":"2606.19643","paper_version":1,"verdict":"UNVERDICTED","confidence":"LOW","novelty_score":6.0,"correctness_risk":"unknown","formal_verification":"none","parameter_count":0,"one_line_summary":"Extends variational consensus Monte Carlo to over-fitted Bayesian mixture models for federated settings, adding cluster-matching algorithms and inference strategies for cases where clusters may not appear in every silo.","lead":"The paper develops a federated learning pipeline for Bayesian mixture models using an extended variational consensus Monte Carlo method that handles unknown numbers of clusters without requiring conjugacy. This could enable privacy-preserving analysis of sensitive health data across institutions without pooling raw records.","discovery_kind":"new_method","skeptic_critique":{"model":"grok-4.3","headline":"Variational aggregation may fail to accurately approximate the global posterior for over-fitted mixtures when cluster presence varies across silos","rationale":"The reader's weakest assumption exactly identifies the variational aggregation step for over-fitted non-conjugate mixtures with varying cluster presence. This matches the load-bearing point for the strongest claim about small-cluster recovery. No stronger internal inconsistency or unsupported derivation is evident from the provided abstract and claim structure.","tokens_in":1792,"tokens_out":301,"duration_ms":19454,"concrete_test":"In the simulation study, disable or replace the novel cluster-matching step with a baseline (e.g., Hungarian algorithm on component means only) and recompute the small-cluster recovery metrics; if the advantage over pooled MCMC vanishes while other factors remain fixed, the matching/aggregation step is the load-bearing element.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim requires that local MCMC posteriors (from independent silo runs) can be variationally aggregated to recover the pooled posterior, including accurate probabilities for small clusters. This holds only if the novel cluster-matching algorithms correctly align components despite non-conjugacy and partial cluster overlap. The aggregation is framed as a variational problem extending Rabinovich et al. (2015), but without closed-form local posteriors the optimization must simultaneously resolve label switching and missing clusters; any systematic mismatch in the matching step would distort the aggregated weights for rare components, undermining the reported accuracy gain over pooled MCMC.","agreement_with_reader":"agree"},"referee_report":{"model":"grok-4.3","summary":"The manuscript presents a variational extension of Consensus Monte Carlo for Bayesian mixture models in federated settings where data cannot be pooled. It extends the approach of Rabinovich et al. (2015) to over-fitted mixtures that infer the number of clusters without requiring conjugacy, introduces novel cluster-matching algorithms to handle cases where clusters are absent from some local datasets, provides multiple inference strategies for the aggregation step, and includes guidelines for their use. A simulation study is used to validate the framework and compare against federated alternatives, with the key result that the method recovers small clusters more accurately than standard MCMC on pooled data when local dataset composition reflects the underlying structure; the approach is illustrated on large-scale electronic health record data for multi-morbidity patterns.","tokens_in":1915,"tokens_out":551,"duration_ms":17715,"significance":"If the variational aggregation and cluster-matching steps accurately recover the global posterior, including for rare components, the work would be significant for privacy-preserving Bayesian inference on mixture models, particularly in health applications. The extension to over-fitted non-conjugate mixtures and the handling of partial cluster overlap across silos address practical federated constraints. The simulation study and real-data illustration provide concrete evidence of utility, though the strength of the accuracy claims depends on the robustness of the reported comparisons.","major_comments":[{"comment":"Simulation study section: the central claim that the method recovers small clusters with greater accuracy than pooled MCMC when local data composition reflects clustering structure is not accompanied by details on error bars, exact cluster-matching algorithms employed, data exclusion rules, or quantitative metrics and baselines used in the comparisons; without these, the reported performance gain cannot be evaluated.","section":"Simulation study"},{"comment":"Section on cluster-matching algorithms (extending Rabinovich et al.): the variational aggregation for over-fitted mixtures requires simultaneous resolution of label switching and missing clusters across silos without closed-form local posteriors; the manuscript does not specify the optimization procedure or provide evidence that systematic mismatches in matching do not distort aggregated weights for rare components, which is load-bearing for the accuracy claim over pooled MCMC.","section":"Cluster-matching algorithms"}],"minor_comments":[{"comment":"The reference to Rabinovich, Angelino and Jordan (2015) in the abstract and introduction should be expanded with full bibliographic details in the reference list.","section":"Introduction"},{"comment":"Notation for the variational objective and cluster alignment parameters could be clarified with an explicit table or diagram to aid reproducibility of the aggregation step.","section":null}],"recommendation":"major_revision","confidential_remarks":null},"author_rebuttal":{"model":"grok-4.3","summary":"We thank the referee for their detailed and constructive review. We respond to each major comment below and indicate the revisions we will make to address them.","responses":[{"response":"We acknowledge that the simulation study section would benefit from greater detail to allow full evaluation of the performance claims. In the revised manuscript, we will include error bars for all metrics, provide the exact specifications and pseudocode for the cluster-matching algorithms, clarify any data exclusion rules, and expand on the quantitative metrics and baselines with additional tables or text. This will make the comparisons more transparent and reproducible.","revision_made":"yes","referee_comment":"[Simulation study] Simulation study section: the central claim that the method recovers small clusters with greater accuracy than pooled MCMC when local data composition reflects clustering structure is not accompanied by details on error bars, exact cluster-matching algorithms employed, data exclusion rules, or quantitative metrics and baselines used in the comparisons; without these, the reported performance gain cannot be evaluated."},{"response":"The cluster-matching algorithms are detailed in the manuscript as extensions to handle missing clusters and label switching in the variational framework. To address the concern about the optimization procedure, we will add a more explicit description of the variational optimization steps, including the objective function and the algorithm used for solving it. Furthermore, we will include supplementary analysis or additional simulation results to demonstrate that any potential mismatches do not systematically bias the weights of rare components, thereby supporting the accuracy claims.","revision_made":"yes","referee_comment":"[Cluster-matching algorithms] Section on cluster-matching algorithms (extending Rabinovich et al.): the variational aggregation for over-fitted mixtures requires simultaneous resolution of label switching and missing clusters across silos without closed-form local posteriors; the manuscript does not specify the optimization procedure or provide evidence that systematic mismatches in matching do not distort aggregated weights for rare components, which is load-bearing for the accuracy claim over pooled MCMC."}],"tokens_in":1518,"tokens_out":417,"duration_ms":30503,"standing_objections":[]},"desk_editor":{"model":"grok-4.3","letter":"The main advance here is taking the 2015 variational consensus Monte Carlo method and extending it to over-fitted Bayesian mixtures that infer the number of clusters, without conjugacy, plus new algorithms to match clusters when some silos lack certain components. They also lay out several aggregation strategies matched to different federated constraints and give guidelines for picking among them. A simulation study and an EHR application on multi-morbidity patterns round it out.\n\nThat pipeline is useful for health-data settings where pooling is blocked by privacy rules. The claim that the method recovers small clusters more accurately than pooled MCMC when local data composition reflects the global structure is the part that stands out if it holds.\n\nThe soft spot is that the abstract gives no error bars, no quantitative tables on matching error, and no clear description of how the variational step resolves label switching plus missing clusters at the same time. Without those, it is difficult to judge whether the aggregation really approximates the global posterior better than the obvious alternative, especially for rare components. The stress-test concern about systematic mismatch in the matching step looks like it could bite if the optimization is not robust.\n\nThis is for people working on federated Bayesian nonparametrics or applied mixture modeling in sensitive domains. A reader who needs a concrete federated pipeline for mixtures would get practical value from the strategies and guidelines.\n\nIt deserves a serious referee because the problem is real, the extensions are concrete, and the simulation plus real-data illustration give something to evaluate even if revisions are needed on the matching and comparison details.","headline":"The paper extends variational CMC to over-fitted mixtures with new cross-silo matching, but the accuracy claim over pooled MCMC rests on details that need checking.","tokens_in":2423,"tokens_out":386,"would_cite":false,"duration_ms":24861,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"grok-4.3","headline":"Variational aggregation of local MCMC posteriors approximates the global posterior for over-fitted Bayesian mixtures even when clusters are absent from some silos.","keywords":["Bayesian mixture models","federated learning","consensus Monte Carlo","variational inference","over-fitted mixtures","cluster matching","electronic health records","multi-morbidity"],"falsifier":"Apply the method and pooled MCMC to simulated data containing known small clusters that appear in only some silos, then compare the estimated posterior probabilities or cluster assignments for those small clusters against ground truth.","tokens_in":2692,"feed_emoji":"","tokens_out":652,"duration_ms":31091,"temperature":0.7,"pith_summary":"The paper develops a method for Bayesian mixture inference when data stay in separate silos due to privacy constraints. Independent MCMC runs produce local posteriors that are then variationally aggregated, with new cluster-matching steps to handle mixtures where not every cluster appears in every silo and where the number of clusters is unknown. Simulations demonstrate that this recovers small clusters more accurately than standard MCMC on the full pooled data when the local datasets reflect the overall clustering structure. The approach is illustrated on electronic health records for identifying multi-morbidity patterns.","feed_headline":"Silo MCMC beats pooled MCMC at finding small clusters in mixtures","feed_subtitle":"When local dataset compositions reflect global structure, variational aggregation recovers rare clusters more accurately without centralizin","key_machinery":"Variational consensus Monte Carlo aggregation of local posteriors, extended by cluster-matching algorithms that align components across silos with incomplete cluster presence.","core_discovery":"Local posterior distributions obtained from independent MCMC runs on each silo can be variationally aggregated to accurately approximate the global posterior over the full pooled data for over-fitted Bayesian mixture models that infer the number of clusters and all parameters without requiring conjugacy, using novel cluster-matching algorithms suitable for settings in which not every cluster appears in each local dataset; when the composition of local datasets reflects the underlying clustering structure, this recovers small clusters with greater accuracy than standard MCMC applied to the pooled data.","pith_inferences":["The same aggregation logic could be tested on other model classes where local MCMC is feasible but full pooling is not.","Accuracy is expected to degrade when clusters are randomly rather than structurally distributed across silos, providing a clear comparison point.","The approach suggests a route to federated clustering analyses in regulated domains such as finance or genomics."],"forward_implications":["Inference proceeds without ever pooling or sharing the raw data across silos.","The framework applies to over-fitted mixtures that determine the number of clusters from the data.","Multiple aggregation strategies are provided and can be selected according to specific federated constraints.","Small clusters are recovered more accurately than pooled MCMC when silo compositions mirror the global structure."],"fun_headline_variants":["Silo variational CMC outperforms pooled MCMC on small mixture clusters","Variational aggregation of silo MCMCs recovers small clusters better than pooled","Federated variational CMC detects small clusters more accurately than pooled MCMC","Bayesian mixtures in silos use variational CMC to recover rare clusters accurately"],"cache_read_input_tokens":2112,"weakest_assumption_plain":"Local posteriors computed separately in each silo can be variationally combined into an accurate approximation of the single global posterior even for non-conjugate over-fitted mixtures and uneven cluster distribution across silos.","fun_headline_variants_meta":{"raw":{"variants":["Silo variational CMC outperforms pooled MCMC on small mixture clusters","Variational aggregation of silo MCMCs recovers small clusters better than pooled","Federated variational CMC detects small clusters more accurately than pooled MCMC","Bayesian mixtures in silos use variational CMC to recover rare clusters accurately"]},"model":"grok-4.3","cost_usd":0.006269,"raw_usage":{"total_tokens":2909,"prompt_tokens":749,"num_sources_used":0,"completion_tokens":70,"cost_in_usd_ticks":62690500,"prompt_tokens_details":{"text_tokens":749,"audio_tokens":0,"image_tokens":0,"cached_tokens":64},"completion_tokens_details":{"audio_tokens":0,"reasoning_tokens":2090,"accepted_prediction_tokens":0,"rejected_prediction_tokens":0}},"tokens_in":749,"tokens_out":70,"duration_ms":16149,"temperature":1.0,"reasoning_tokens":2090,"cache_read_input_tokens":64,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-06-26T18:38:15.045844+00:00","model_set":{"reader":"grok-4.3"},"falsifier":"Apply the method and pooled MCMC to simulated data containing known small clusters that appear in only some silos, then compare the estimated posterior probabilities or cluster assignments for those small clusters against ground truth.","supporting_citations":[],"review_version":1}